release CUDA memory in WeightConverter and avoid meaningless print

[xin3he]

Description

release CUDA memory in WeightConverter and avoid meaningless print

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please specify):

Related Issues

Fixes or relates to #1497

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.

[Copilot]

Pull request overview

This PR addresses a VRAM memory leak during weight conversion in the WeightConverter (issue #1497), where CUDA memory was accumulating across layers because intermediate tensors and original layer data were not being freed. It also refines memory monitoring by moving per-module logging into the unfuse function and adjusting a decorator message.

Changes:

• Added del dq_weight and layer.to("meta") in all four convert_layer implementations (FP8, MXFP4, MXFP8, NVFP4) to release intermediate CUDA tensors and original layer data after dequantization.
• Moved memory monitoring from the @dump_mem_usage decorator on _handle_moe_modules into _unfuse_experts_weights_inplace for per-module VRAM tracking.
• Changed the capitalization of the dump_mem_usage message for _apply_custom_replacements.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
auto_round/utils/weight_handler.py	Added del dq_weight and layer.to("meta") in all four handler convert_layer methods to free CUDA memory after weight dequantization.
auto_round/modeling/fused_moe/replace_modules.py	Removed @dump_mem_usage decorator from _handle_moe_modules; changed capitalization on _apply_custom_replacements decorator message.
auto_round/modeling/fused_moe/moe_experts_interface.py	Added memory_monitor.update() and log_summary() calls before/after the unfusing logic in _unfuse_experts_weights_inplace.

---

You can also share your feedback on Copilot code review. Take the survey.